On Tue, 1 Apr 2003, Ville-Pertti Keinonen wrote:

> 
> > For structures 32 words or less, you can do it like (using x86 assembly
> > language here):
> 
>    ...actual code snipped...
> 
> Ok, that's not necessarily too bad.
> 
> I guess I was just intuitively dismissing anything that felt like it 
> had more than a couple of instructions of "overhead".   I'm probably 
> also biased by SMP-thinking - read-modify-write operations feel 
> inherently evil. ;-)
> 

For SMP you want different minor heaps for every thread.  Adding the lock 
prefix to an instruction costs ~100 clock cycles.  Much better to simply 
give each thread it's own minor heap.  64K isn't that much memory.  And 
there are stunts you can pull if you threads spend most of their life 
blocked to decrease the number of minor heaps you need.

Note that you implicitly have a read-modify-write cycle already in your
allocation.  You need to read the top-of-heap pointer, save off the
original pointer, increment by the size you are allocating, and write the 
incremented pointer back, all automatically.  A wonderful application of 
load locked/store conditional, if only we had it in the x86.  Without it, 
you basically need to grab a spinlock before updating.  Thus the lock 
prefix.

Brian


-------------------
To unsubscribe, mail caml-list-request@inria.fr Archives: http://caml.inria.fr
Bug reports: http://caml.inria.fr/bin/caml-bugs FAQ: http://caml.inria.fr/FAQ/
Beginner's list: http://groups.yahoo.com/group/ocaml_beginners